First, you need to download the Jetty servlet container for your platform. You can get your download package from an automatic installer, such as apt-get, or you can download it from http://download.eclipse.org/jetty/. In addition to this, read the Using core discovery recipe of this chapter for more information.

The first step is to install the Jetty servlet container, which is beyond the scope of this book, so we will assume that you have Jetty installed in the /usr/share/jetty directory.

We can now run Jetty to see if everything is okay. To start Jetty, which was already installed, use the apt-get command, as shown:

/etc/init.d/jetty start

If there are no exceptions during startup, we have a running Jetty with Solr deployed and configured. To check whether Solr is running, visit http://localhost:8983/solr/.

Congratulations, you have just successfully installed, configured, and run the Jetty servlet container with Solr deployed.

For the purpose of this recipe, I assumed that we needed a single core installation with only the schema.xml and solrconfig.xml configuration files. Multicore installation is very similar; it differs only in terms of the Solr configuration files—one needs more than a single core defined.

The first thing we did was copied the solr.war file and created the tmp directory. The WAR file is the actual Solr web application. The tmp directory will be used by Jetty to unpack the WAR file.

The solr-jetty-context.xml file that we place in the context directory allows Jetty to define the context for a Solr web application. As you can see in its contents, we have set the context to be /solr, so our Solr application will be available under http://localhost:8983/solr/. We also need to specify where Jetty should look for the WAR file (the war property), where the web application descriptor file (the defaultsDescriptor property) is, and finally, where the temporary directory will be located (the tempDirectory property).

Copying the jetty.xml and webdefault.xml files is important. The standard Solr distribution comes with Jetty configuration files prepared for high load; for example, we can avoid the distributed deadlock.

The next step is to provide configuration files for the Solr core. These files should be put in the core_name/conf directory, which is created in a folder specified by the system's solr.solr.home variable. Since our core is named example_data, and the solr.solr.home property points to /usr/share/solr, we place our configuration files in the /usr/share/solr/example_data/conf directory. Note that I decided to use the /usr/share/solr directory as the base directory for all Solr configuration files. This ensures the ability to update Jetty without the need to override or delete the Solr configuration files.

The core.properties file allows Solr to identify the core that it will try to load. By providing the name property, we tell Solr what name the core should have. In our case, its name will be example_data.

The zoo.cfg file is optional, is only needed when setting up SolrCloud, and is used by Solr to specify ZooKeeper client properties. The tickTime property specifies the number of milliseconds of each tick. The tick is the unit of time in ZooKeeper client connections. The initLimit property specifies the number of ticks the initial synchronization phase can take, and the syncLimit property specifies the number of ticks that can pass between sending a request and getting an acknowledgement. For example, because the syncLimit property is set to 5 and tickTime is 2000, the maximum time between sending the request and getting the acknowledgement is 10,000 milliseconds (syncLimit multiplied by tickTime).

The solr.xml file is described in the Using core discovery recipe in this chapter.

If you installed Jetty with the apt-get command or a similar software, then you need to update the /etc/default/jetty file to include the solr.solr.home variable for Solr to be able to see its configuration directory.

After all these steps, we will be ready to launch Jetty. If you installed Jetty with apt-get or similar software, you can run Jetty with the first command shown in the example. Otherwise, you can run Jetty with the java -jar start command from the Jetty installation directory.

After running the example query in your web browser, you should see the Solr front page as a single core. Congratulations, you have successfully configured and run the Jetty servlet container with Solr deployed.

There are a few more tasks that you can perform to counter some problems while running Solr within the Jetty servlet container. The most common tasks that I encountered during my work are described in the ensuing sections.

<Set name="requestHeaderSize">32768</Set>

<Call name="addConnector">
 <Arg>
  <New class="org.mortbay.jetty.bio.SocketConnector">
   <Set name="port"><SystemProperty name="jetty.port"  
      default="8080"/></Set>
   <Set name="maxIdleTime">50000</Set>
   <Set name="lowResourceMaxIdleTime">1500</Set>
   <Set name="requestHeaderSize">32768</Set>
  </New>
 </Arg>
</Call>

The installation instructions in this recipe contain information about installing ZooKeeper Version 3.4.6, but it should be useable for any minor release changes of Apache ZooKeeper. To download ZooKeeper, visit http://zookeeper.apache.org/releases.html. This recipe will show you how to install ZooKeeper in a Linux-based environment. For ZooKeeper to work, Java needs to be installed.

Let's assume that we have decided to install ZooKeeper in the /usr/share/zookeeper directory of our server, and we want to have three servers (with IPs 192.168.1.1, 192.168.1.2, and 192.168.1.3) hosting a distributed ZooKeeper installation. This can be done by performing the following steps:

That's all. Of course, you can also add the ZooKeeper service to start automatically as your operating system starts up, but this is beyond the scope of the recipe and book.

I talked about the ZooKeeper quorum and started this using three ZooKeeper nodes. ZooKeeper operates in a quorum, which means that at least 50 percent plus one server needs to be available and connected. We can start with a single ZooKeeper server, but such deployment won't be highly available and resistant to failures. So, to be able to handle at least a single ZooKeeper node failure, we need at least three ZooKeeper nodes running.

Let's skip the first part because creating the directory and unpacking the ZooKeeper server is quite simple. What I would like to concentrate on are the configuration values of the ZooKeeper server. The clientPort property specifies the port on which our SolrCloud servers should connect to ZooKeeper. The dataDir property specifies the directory where ZooKeeper will hold its data. Note that ZooKeeper needs read and write permissions to the directory. So far so good, right? So, now, the more advanced properties, such as tickTime, specified in milliseconds is the basic time unit for ZooKeeper. The initLimit property specifies how many ticks the initial synchronization phase can take. Finally, syncLimit specifies how many ticks can pass between sending the request and receiving an acknowledgement.

There are also three additional properties present, server.1, server.2, and server.3. These three properties define the addresses of the ZooKeeper instances that will form the quorum. The values for each of these properties are separated by a colon character. The first part is the IP address of the ZooKeeper server, and the second and third parts are the ports used by ZooKeeper instances to communicate with each other.

The last thing is the myid file located in the /usr/share/zookeeper/data directory. The contents of the file is used by ZooKeeper to identify itself. This is why we need to properly configure it so that ZooKeeper is not confused. So, for the ZooKeeper server specified as server.1, we need to write 1 to the myid file.

Before continuing further, it is advised to read the Installing Zookeeper for SolrCloud and Running Solr on a standalone Jetty recipes of this chapter. They will show you how to set up a Zookeeper cluster to be ready for production use and how to configure Jetty and Solr to work with each other.

That's all. Your Solr instance configuration files are now ready to be used with SolrCloud.

Now, let's see why all these changes are needed in order to use our old configuration files with SolrCloud.

The _version_ field is used by Solr to enable document versioning and optimistic locking, which ensures that you won't have the newest version of your document overwritten by mistake. As a result of this, SolrCloud requires the _version_ field to be present in the index structure. Adding this field is simple—you just need to place another field definition that is stored, indexed, and based on a long type, that's all.

As for the replication handler, you should remember not to add slave- or master-specific configurations, but only a simple request handler definition, as shown in the previous example. The same applies to the administration panel handlers; they need to be available under the default URL address.

The real-time get handler is responsible for getting the updated documents right away. In general, the documents are not available to search if the Lucene index searcher is not open, which happens after a hard or soft commit command (we will talk more about commit and soft commit in the Configuring SolrCloud for NRT use cases recipe of this chapter). This handler allows Solr (and also you) to retrieve the latest version of the document without the need to reopen the searcher, and thus, even if the document is not yet visible during a usual search operation. This is done by using the transaction log if the document is not yet indexed. The configuration is very similar to usual request handler configurations; you need to add a new handler with the name property set to /get and the class property set to solr.RealTimeGetHandler. In addition to this, we want the handler to omit response headers (the omitHeader property set to true) and return a response in JSON (with the wt property set to json). We omit the headers so that we have responses that are easier to parse.

One of the last things that is needed by SolrCloud is the transaction log, which enables real-time get operations to be functional. The transaction log keeps track of all the uncommitted changes and enables real-time get handlers to retrieve them. In order to turn on transaction log usage, one should add the updateLog tag to the solrconfig.xml file and specify the directory where the transaction log directory should be created (by adding the dir property, as shown in the example). In the previous configuration, we tell Solr that we want to use the Solr data directory as the place to store transaction log directories.

Finally, Solr needs you to keep the default address for the core administrative interface, so you should remember to have the adminPath property set to the value shown in the example (in the solr.xml file). This is needed in order for Solr to be able to manipulate cores.

We already talked about the solr.xml file contents in the Running Solr on a standalone Jetty recipe in this chapter, so refer to that recipe if you are not familiar with the contents.

In order to use the desired directory, all you need to do is choose the right directory factory implementation and inform Solr about it. Let's assume that you want to use NRTCachingDirectory as your directory implementation. In order to do this, you need to place (or replace if it is already present) the following fragment in your solrconfig.xml file:

<directoryFactory name="DirectoryFactory" class="solr.NRTCachingDirectoryFactory" />

That's all. The setup is quite simple, but I think that the question that will arise is what directory factories are available to use. When this book was written, the following directory factories were available:

Now, let's see what each of these factories provides.

Before we get into the details of each of the presented directory factories, I would like to comment on the directory factory configuration parameter. All you need to remember is that the name attribute of the directoryFactory tag should be set to DirectoryFactory, and the class attribute should be set to the directory factory implementation of your choice. Also, some of the directory implementations can take additional parameters that define their behavior. We will talk about some of them in other recipes in the book (for example, in the Limiting I/O usage recipe in this chapter).

If you want Solr to make decisions for you, you should use the solr.StandardDirectoryFactory directory factory. It is filesystem-based and tries to choose the best implementation based on your current operating system and Java virtual machine used. If you implement a small application that won't use many threads, you can use the solr.SimpleFSDirectoryFactory directory factory that stores the index file on your local filesystem, but it doesn't scale well with a high number of threads. The solr.NIOFSDirectoryFactory directory factory scales well with many threads, but remember that it doesn't work well on Microsoft Windows platforms (it's much slower) because of a JVM bug (http://bugs.java.com/bugdatabase/view_bug.do?bug_id=6265734).

The solr.MMapDirectoryFactory directory factory has been the default directory factory for Solr for 64-bit Linux systems since Solr 3.1. This directory implementation uses virtual memory and the kernel feature called mmap to access index files stored on disk. This allows Lucene (and thus Solr) to directly access the I/O cache. This is desirable, and you should stick to this directory if near real-time searching is not needed.

If you need near real-time indexing and searching, you should use solr.NRTCachingDirectoryFactory. It is designed to store some parts of the index in memory (small chunks), and thus speeds up some near real-time operations greatly. By saying near real-time, we mean that the documents are available within milliseconds from indexing.

The solr.HdfsDirectoryFactory is used when Solr runs on HDFS filesystems, so inside a Hadoop cluster. If you are using Solr inside a Hadoop cluster, then it is almost certain that you'll want to use the directory implementation.

The last directory factory, solr.RAMDirectoryFactory, is the only one that is not persistent. The whole index is stored in the RAM memory, and thus, you'll lose your index after a restart or server crash. Also, you should remember that replication won't work when using solr.RAMDirectoryFactory. One might ask why I should use this factory? Imagine a volatile index autocomplete functionality or for unit tests of your query's relevance, or just anything you can think of when you don't need to have persistent and replicated data. However, remember that this directory is not designed to hold large amounts of data.

First, let's assume we have a field in the index called title in which we hold the titles of our documents. What's more, we don't want the spellchecker to have its own index, and we would like to use this title field to provide spelling suggestions. In addition, we would like to decide when we want spelling suggestions. In order to do this, we need to do two things:

If you check your data folder, you will see that there is no directory responsible for holding the spellchecker index. Now, let's see how this works.

Now, let's get into some specifics about how the configuration shown in the preceding example works. We will start from the search component configuration. The queryAnalyzerFieldType property tells Solr which field configuration should be used to analyze the query passed to the spellchecker. The name property sets the name of the spellchecker, which is used in the handler configuration later. The field property specifies which field should be used as the source for the data used to build spelling suggestions. As you probably figured out, the classname property specifies the implementation class, which in our case is solr.DirectSolrSpellChecker, enabling us to omit having a separate spellchecker index; spellchecker will just use the main index. The next parameters visible in the previous configuration specify how the Solr spellchecker should behave; however, this is beyond the scope of this recipe (if you want to read more about the parameters, visit the http://wiki.apache.org/solr/SpellCheckComponent URL).

The last thing is the request handler configuration. Let's concentrate on all the properties that start with the spellcheck prefix. First, we have spellcheck.dictionary, which, in our case, specifies the name of the spellchecking component we want to use (note that the value of the property matches the value of the name property in the search component configuration). We tell Solr that we want spellchecking results to be present (the spellcheck property with the on value), and we also tell Solr that we want to see the extended result format, which allows us to see more with regard to the results (spellcheck.extendedResults set to true). In addition to the previous configuration properties, we also said that we want to have a maximum of five suggestions (the spellcheck.count property), and we want to see the collation and its extended results (spellcheck.collate and spellcheck.collateExtendedResults both set to true).

Let's see one more thing—the ability to have more than one spellchecker defined in a request handler.

<requestHandler name="/spell" class="solr.SearchHandler" startup="lazy">
<requestHandler name="/spell" class="solr.SearchHandler" startup="lazy">
 <lst name="defaults">
  <str name="df">title</str>
  <str name="spellcheck.dictionary">direct</str>
  <str name="spellcheck.dictionary">word</str>
  <str name="spellcheck.dictionary">better</str>
  <str name="spellcheck">on</str>
  <str name="spellcheck.extendedResults">true</str>
  <str name="spellcheck.count">5</str>
  <str name="spellcheck.collate">true</str>
  <str name="spellcheck.collateExtendedResults">true</str>
 </lst>
 <arr name="last-components">
  <str>spellcheck</str>
 </arr>
</requestHandler>

Let's assume that we don't know anything about the data and we want to fully rely on Solr when it comes to it.

We start with our index having two fields, id and _version_. The id field is used as the unique identifier; we informed Solr about this by adding the unqiueKey section in schema.xml. We will need it for functionalities such as document updates, deletes by identifiers, and so forth. The _version_ field is used by Solr internally, and is required by some Solr functionalities (such as optimistic locking); this is why we include it. The rest of the fields will be added automatically.

We also need to define the field types that we will use. Apart from the string type used by the id field, and the long type used by the _version_ field, it contains types our documents will use. We will also define these types in our custom processor chain in the solrconfig.xml file.

The next thing is very important; the managed schema factory that we defined in solrconfig.xml, which is a ManagedIndexSchemaFactory type (the class property set to this value). By adding this section, we say that we want Solr to manage our schema.xml file. This means that Solr will load the schema.xml file during startup, change its name to schema.xml.bak, and will then create a file called managed-schema (the value of the managedSchemaResourceName property). From this point, we shouldn't modify our index structure manually—we should either let Solr do it during indexation or add and alter fields using the schema API (we will talk about this in the Altering the index structure on a live collection recipe in Chapter 8, Using Additional Functionalities). Since I assume that we will use the schema API, I've set the mutable property to true. If we want to disallow using the schema API, we should set the mutable property to false.

We also need to include an update request processor chain. Since we want all index requests to use our custom request chain, we add the update.chain property and set it to add-unknown-fields in the defaults section of our update request handler configuration.

Finally, the second most important thing in this recipe is our update request processor chain called add-unknown-fields (the same as we used in the update processor configuration). It defines several update processors that allow us to get the functionality of fields and their types' discoveries. The solr.RemoveBlankFieldUpdateProcessorFactory processor factory removes empty fields from the documents we send to indexation. The solr.ParseBooleanFieldUpdateProcessorFactory processor factory is responsible for parsing Boolean fields; solr.ParseLongFieldUpdateProcessorFactory parses fields that have data that uses the long type; solr.ParseDoubleFieldUpdateProcessorFactory parses fields with data of double type; and solr.ParseDateFieldUpdateProcessorFactory parses the date-based fields. We specify the format we want Solr to recognize (we will discuss this in more detail in the Using parsing update processors to parse data recipe in Chapter 2, Indexing Your Data).

Finally, we include the solr.AddSchemaFieldsUpdateProcessorFactory processor factory that adds the actual fields to our managed schema. We specify the default field type to text by adding the defaultFieldType property. This type will be used when no other type will match the field. After the default field type definition, we see four lists called typeMapping. These sections define the field type mappings Solr will use. Each list contains at least one valueClass property and one fieldType property. The valueClass property defines the type of data Solr will assign to the field type defined by the fieldType property.

In our case, if Solr finds a date (<str name="valueClass">java.util.Date</str>) value in a field, it will create a new field using the tdates field type (<str name="fieldType">tdates</str>). If Solr finds a long or an integer value, it creates a new field using the tlongs field type. Of course, a field won't be created if it already exists in our managed schema. The name of the field created in our managed schema will be the same as the name of the field in the indexed document.

Finally, the solr.LogUpdateProcessorFactory processor factory tells Solr to write information about the update to log, and the solr.RunUpdateProcessorFactory processor factory tells Solr to run the update itself.

As we can see, our data includes fields that we didn't specify in the schema.xml file, and the document was indexed properly, which allows us to assume that the functionality works. If you want to check how our index structure looks like after indexation, use the schema API; you can do it yourself after reading the Retrieving information about the index structure recipe in Chapter 8, Using Additional Functionalities.

One thing to remember is that by default, Solr is able to automatically detect field types such as Boolean, integer, float, long, double, and date.

Before continuing further with this recipe, read the Choosing the proper directory configuration recipe of this chapter to see what directories are available and how to configure them.

Let's assume that we want to limit the I/O usage for our use case that uses solr.MMapDirectoryFactory. So, in the solrconfig.xml file, we will have the following configuration present:

<directoryFactory name="DirectoryFactory" class="solr.MMapDirectoryFactory">
</directoryFactory>

Now, let's introduce the following limits:

To do this, we change our previous configuration to the following:

<directoryFactory name="DirectoryFactory" class="solr.MMapDirectoryFactory">
 <double name="maxWriteMBPerSecFlush">20</double>
 <double name="maxWriteMBPerSecMerge">10</double>
 <double name="maxWriteMBPerSecRead">50</double>
</directoryFactory>

After altering the configuration, all we need to do is restart Solr and the limits will be taken into consideration.

The logic behind setting the limits is very simple. All directories that extend the Solr CachingDirectoryFactory class allow us to set the maxWriteMBPerSecFlush, maxWriteMBPerSecMerge and maxWriteMBPerSecRead properties. The mentioned directory implementations are all the directory implementations that were mentioned in the Choosing the proper directory configuration recipe of this chapter.

The maxWriteMBPerSecFlush property allows us to tell Solr how many megabytes per second can be written by Solr during segment flush (so, during the write operation that is not triggered by segment merging). The maxWriteMBPerSecMerge property allows us to specify how many megabytes per second can be written by Solr during segment merge. Finally, the maxWriteMBPerSecRead property specifies the amount of megabytes allowed to be read per second. One thing to remember is that the values are approximated, not exact.

Limiting I/O usage can be very handy, especially in deployments where I/O usage is at its maximum. During query peak hours, when we want to solve server queries as fast as we can, we need to minimize the indexing and merging impact. With proper configuration that is adjusted to our needs, we can just limit the I/O usage and still serve queries with the latency we want.

Using the new core discovery process is very simple.

That's all; during startup, Solr will load our core.

The solr.xml file is the same one that is provided with the Solr example deployment, and it contains the default values related to Solr configuration. The host property specifies the hostname, and the hostPort property specifies the port on which Solr will run (it will be taken from the jetty.port property, and is by default 8983). The hostContext property specifies the web application context under which Solr will be available (by default, it is solr). In addition to this, we can specify the ZooKeeper client session timeout by using the zkClientTimeout property (used only in the SolrCloud mode, defaulting to 30,000 milliseconds). By default, we also say that we want Solr to use generic core names for SolrCloud, and we can change this by specifying false in the genericCoreNodeNames property.

There are two additional properties that relate to shard handling. The socketTimeout property specifies the timeout of socket connection, and the connTimeout property specifies the timeout of connection. Both the properties are used to create clients used by Solr to communicate between shards. The connection timeout specifies the timeout when Solr connects to another shard, and it takes a long time; the socket timeout is about the time to wait for the response to be back.

The simplest core.properties file is an empty file, in which case, Solr will try to choose the core name for us. However, in our case, we wanted to give the core a name we've chosen, and because of this, we included a single name entry that defines the name Solr will assign to the core. You should remember that Solr will try to load all the cores that have the core.properties file present, and the core name doesn't have to live in the directory of the same name.

Of course, the name property is not the only property available for usage. There are other properties, but in most cases, you'll use the name property only:

Finally, it is worth saying that the old solr.xml format will not be supported in Solr 5.0, so it is good to get familiar with the new format now.

If you want to see all the properties and sections exposed by the new solr.xml format, refer to the official Apache Solr documentation located at https://cwiki.apache.org/confluence/display/solr/Format+of+solr.xml.

I assume that you already have SolrCloud set up and ready to go (if you don't, refer to the Creating a new SolrCloud cluster recipe in Chapter 7, In the Cloud); you will now know how to update your collection configuration and be interested in near real-time search.

Let's assume that we want our data to be available about one second after it's indexed. To do this, we need to change the solrconfig.xml file so that its update handler section looks as shown:

<updateHandler class="solr.DirectUpdateHandler2">
 <updateLog>
  <str name="dir">${solr.ulog.dir:}</str>
 </updateLog>
 
 <autoSoftCommit>
  <maxTime>1000</maxTime>
 </autoSoftCommit>
 
 <autoCommit>
  <maxTime>300000</maxTime>
  <openSearcher>false</openSearcher>
 </autoCommit>
</updateHandler>

That's all; after a restart or configuration reload, documents should be available to search after about one second.

By changing the configuration of the update handler, we introduced three things. First, using the <updateLog> section, we told Solr to use the update log functionality. The transaction log (another name for this functionality) is a file where Solr writes raw documents so that they can be used in a recovery process. In SolrCloud, each instance of Solr needs to have its own transaction log configured. When a document is sent for indexation, it gets forwarded to the shard leader and the leader sends the document to all its replicas. After all the replicas respond to the leader, the leader itself responds to the node that sent the original request, and this node reports the indexing status to the client. At this point in time, the document is written into a transaction log, not yet indexed, but safely written; so, if a failure occurs (for example, the server shuts down), the document is not lost. During a startup process, the transaction log is replayed and the documents stored in it are indexed, so even if they were not indexed, they will be if a failure happens. After the process of storing the data in transaction logs, Solr can easily index the data located there.

The second thing is the autoSoftCommit section. This is a new autocommit option introduced in Solr 4.0. It basically allows us to reopen the index searcher without closing and opening a new one. For us, this means that our documents that were sent for indexation will start to be visible and available to search. We do this once every 1000 milliseconds as configured using the maxTime tag. The soft commit was introduced because reopening is easier to do and is less resource intensive than closing and opening a new index searcher. In addition to this, it doesn't persist the data to disk by creating a new segment.

However, one has to remember that even though the soft commit is less resource intensive, it is still not free. Some Solr caches will have to be reloaded, such as the filter, document, or query result caches. We will get into more configuration details in the Configuring SolrCloud for high-indexing use cases and Configuring SolrCloud for high-querying use cases recipes in this chapter.

The last thing is the autocommit defined in the autoCommit section, which is called the hard autocommit. It is responsible for flushing data and closing the index segment used for it (because of this segment, merge might start in the background). In addition to this, the hard autocommit also closes the transaction log and opens a new one. We've configured this operation to happen every 5 minutes (300000 milliseconds). What we also included is the <openSearcher>false</openSearcher> section. This means that Solr won't open a new index searcher during a hard auto commit operation. We do this on purpose; we define index searcher opening periods in the soft autocommit section. If we set the openSearcher section to true, Solr will close the old index searcher, open a new one, and automatically warm caches. Before Solr 4.0, this was the only way to have documents visible for searching when using autocommit.

One additional thing to remember is that with soft autocommit set to reopen the searcher very often, all the top level caches, such as the filter, document, and query result caches, will be invalidated. It is worth thinking and doing performance tests if the cache (all or some of them) are actually worth being used at all. I would like to give a clear advice here, but this is highly dependent on the use case. You can read more about cache configuration in the Configuring the document cache, Configuring the query result cache, and Configuring the filter cache recipes in Chapter 6, Improving Solr Performance.

Before continuing reading the recipe, read the Running Solr on a standalone Jetty and Configuring SolrCloud for NRT use cases recipes in this chapter.

In very high indexing use cases, there are chances that you'll use bulk indexing to index your data. In addition to this, because we are talking about SolrCloud, we'll use autocommit so that we can leave the data durability and visibility management to Solr. Let's discuss how to prepare configuration for a use case where indexing is high, but the querying is quite low; for example, when using Solr for log centralization solutions.

Let's assume that we are indexing more than 1,000 documents per second and that we have four nodes, each of 12 cores and 64 GB of RAM. Note that this specification is not something we need to index the number of documents, but they are here for reference.

Now, let's discuss what each of these changes mean.

We started with tuning the autocommit setting, which you should be aware of after reading this recipe. Since we are not worried about documents being visible as soon as they are indexed, we set the soft autocommit's maxTime property to 600000. This means that we will reopen the searcher every 10 minutes, so our documents will be visible maximum 10 minutes after they are sent to indexation.

The one thing to look at is the short time for hard commit, which is every 15 seconds (the maxTime property of the autoCommit section set to 15000). We did this because we don't want transaction logs to contain a high number of entries because this can cause problems during the recovery process.

We also increased the default number of threads an index writer can use from the default 8 to 10 by setting the maxIndexingThreads property. Since we have 12 cores on each machine, and we are not querying much, we can allow more threads using the index writer. If the index writer uses the number of threads that's equal to the maxIndexingThreads property, the next thread will wait for one of the currently running to end. Remember that the maxIndexingThreads property sets the maximum allowed indexing threads, which doesn't mean they will be used every time.

We also increased the default RAM buffer size from 100 to 128 using the ramBufferSizeMB property. We did this to allow Lucene to buffer as many documents as needed in memory. If the size of the documents in the buffer is larger than the given value of the ramBufferSizeMB property, Lucene will flush the data to the directory, which will decide what else to do. We have to remember though that we are also using autocommit, so the data will be flushed every 15 seconds because of hard autocommit settings.

In addition to this, you might want to look at the merging policy and segment merge processes as this can become a major bottleneck. If you are interested, refer to the Tuning segment merging recipe in Chapter 9, Dealing with Problems.

Before continuing to read this recipe, read the Running Solr on a standalone Jetty, Configuring SolrCloud for NRT use cases, and Configuring SolrCloud for high-indexing use cases recipes of this chapter.

Giving general advice for high-querying use cases is pretty hard because it very much depends on the data, cluster structure, query structure, and target latency. In this recipe, we will look at three things—configuration, scaling, and overall general advices. Let's assume that we have four nodes, each having 128 GB of RAM and large disks, and we have 100 million documents we want to search across.

We should start with sizing our cluster. In general, this means choosing the right number of nodes, the right number of shards and replicas for your collections, and the memory. The general advice is to index some portion of your data and see how much space is used. For example, assuming you've indexed 1,000 documents and they are taking 1 MB of disk space, we can now calculate the disk space needed by 100 million documents; this will give us about 100 GB of total disk space used. With a replication factor of 2, we will need 200 GB, which means our four nodes should be enough to have the data cached by the operating system. In addition to this, we will need memory for Solr to operate (we can help ourselves calculate how much we will need using http://svn.apache.org/repos/asf/lucene/dev/trunk/dev-tools/size-estimator-lucene-solr.xls).

Given these facts, we can end up with a minimum of four shards and a replication factor of 2, which will give us a leader shard and its replica for each of the four initial shards we created the collection with. However, going for more initial shards might be better for scaling in the later stage of your application life cycle.

After we know some information, we can prepare the autocommit settings. To do this, we alter our solrconfig.xml configuration file and include the following update handler configuration:

<updateHandler class="solr.DirectUpdateHandler2">

 <updateLog>
  <str name="dir">${solr.ulog.dir:}</str>
 </updateLog>

 <autoSoftCommit>
  <maxTime>30000</maxTime>
 </autoSoftCommit>

 <autoCommit>
  <maxTime>600000</maxTime>
  <openSearcher>false</openSearcher>
 </autoCommit>
</updateHandler>

In addition to this, we should adjust caching, which is covered in the Configuring the document cache, Configuring the query result cache, and Configuring the filter cache recipes in Chapter 6, Improving Solr Performance.

In addition to all this, you might want to look at the merging policy and segment merge processes as this can become a major bottleneck. If you are interested, refer to the Tuning segment merging recipe in Chapter 9, Dealing with Problems.

We started with sizing questions and estimations. Remember that the numbers you will extrapolate from the small portion of data are not exact numbers, they are estimations. What's more, we now know that in order to have our index fully cached by the operating system, we will need at least 200 GB of RAM memory that can be used for the system cache because we will have at least one shard and its physical copy. Of course, the four nodes with 128 GB of RAM are more or less a perfect case when we will be able to have our indices cached. This is because we will have a total of 512 GB of RAM across all nodes. Given the fact that we will end up with four leader shards, one on each machine, four replicas, again one on each machine, and that our index will be evenly divided, it will give us 50 GB of data on each node (25 GB for leader and the same for replica because it is an exact copy).

A few words about having more shards—sometimes, if you expect your data to grow, it is good to create a collection with more shards initially and place multiple ones on a single node. This gives more flexibility when you add new nodes; you can migrate some shards without the need to split them, or you can create a new collection with new shards and reindex your data.

Next, we adjust the autocommit section. Since we don't need near real-time searching, we decide not to stress Solr too much and set the soft autocommit to 60000 milliseconds, which means that the data will be visible after 1 minute from indexing. In general, if you will, the more often you reopen the searcher, the more pressure is put on Solr, and thus, the queries will be slower. So, if you query heavily, you should set the soft autocommit to the maximum time allowed by your use case.

Of course, we also included the hard autocommit and set it to be executed every 10 minutes. We decided to go for this because we don't index much data, so the index shouldn't be changed too often, and the transaction log shouldn't be too large.

Setting up the heartbeat mechanism in Solr is very easy. One just needs to add the following section to the solrconfig.xml file:

<requestHandler name="/admin/ping" class="solr.PingRequestHandler">
 <lst name="invariants">
  <str name="q">solrpingquery</str>
 </lst>
</requestHandler>

This is all. Of course, if we need all our cores and collections to respond to the health requests, we should include the previous section in the solrconfig.xml files for all of them. After this, run a query to the admin/ping handler of our Solr instance, for example:

curl 'localhost:8983/solr/heartbeat_core/admin/ping'

Solr will respond with a status response, for example:

<?xml version="1.0" encoding="UTF-8"?>
<response>
<lst name="responseHeader"><int name="status">0</int><int name="QTime">6</int><lst name="params"/></lst><str name="status">OK</str>
</response>

The configuration is really simple; we defined a new request handler that will be available under the /admin/ping address (of course, we have to prefix it with the host address and core name). The class implementing the handle is the one dedicated to handle the heartbeat mechanism request, solr.PingRequestHandler. We also defined that the q parameter for all the ping requests will be solrpingquery and the request won't be able to overwrite this parameter (because we included it in the invariants section). The ping query should be as simple as it can get so that it runs blazingly fast; what's more, it is usually good for it not to return any search results.

As you can see, the response contains the status section, which in our case has the value of OK. In the case of an error, the status section will contain the error code.

The solr.PingRequestHandler handler allows us to enable and disable the heartbeat mechanism without shutting down the whole Solr instance.

<requestHandler name="/admin/ping" class="solr.PingRequestHandler">
 <lst name="invariants">
  <str name="q">solrpingquery</str>
 </lst>
 <str name="healthcheckFile">server-enabled.txt</str>
</requestHandler>

curl 'localhost:8983/solr/heartbeat_core/admin/ping?action=enable'

curl 'localhost:8983/
solr/heartbeat_core/admin/ping?action=disable'

curl 'localhost:8983/solr/heartbeat_core/admin/ping?action=status'

Before choosing one of the score calculation algorithms available in Solr, it's good to read a bit about them. The detailed description of all the algorithms is beyond the scope of this recipe and the book (although a simple description is mentioned later in the recipe), but I suggest visiting the Solr wiki page (or Javadocs) and reading basic information about the available implementations.

For the purpose of this recipe, let's assume we have the following index structure (just add the following entries to your schema.xml file):

<field name="id" type="string" indexed="true" stored="true" required="true" />
<field name="name" type="text_general" indexed="true" stored="true"/>
<field name="description" type="text_general_dfr" indexed="true" stored="true" />

The string and text_general types are available in the default schema.xml file provided with the example Solr distribution. However, we want DFRSimilarity to be used to calculate the score for the description field. In order to do this, we introduce a new type, which is defined as follows (just add the following entries to your schema.xml file):

<fieldType name="text_general_dfr" class="solr.TextField" positionIncrementGap="100">
 <analyzer type="index">
  <tokenizer class="solr.StandardTokenizerFactory"/>
  <filter class="solr.StopFilterFactory" ignoreCase="true" words="stopwords.txt" enablePositionIncrements="true" />
  <filter class="solr.LowerCaseFilterFactory"/>
 </analyzer>
 <analyzer type="query">
  <tokenizer class="solr.StandardTokenizerFactory"/>
  <filter class="solr.StopFilterFactory" ignoreCase="true" words="stopwords.txt" enablePositionIncrements="true" />
  <filter class="solr.SynonymFilterFactory" synonyms="synonyms.txt" ignoreCase="true" expand="true"/>
  <filter class="solr.LowerCaseFilterFactory"/>
 </analyzer>
 <similarity class="solr.DFRSimilarityFactory">
  <str name="basicModel">P</str>
  <str name="afterEffect">L</str>
  <str name="normalization">H2</str>
  <float name="c">7</float>
 </similarity>
</fieldType>

Also, to use the per-field similarity, we have to add the following entry to your schema.xml file:

<similarity class="solr.SchemaSimilarityFactory"/>

That's all. Now, let's have a look and see how this works.

The index structure previously presented is pretty simple as there are only three fields. The one thing we are interested in is that the description field uses our own custom field type called text_generanl_dfr.

The thing we are most interested in is the new field type definition called text_general_dfr. As you can see, apart from the index and query analyzer, there is an additional section called similarity. It is responsible for specifying which similarity implementation to use to calculate the score for a given field. You are probably used to defining field types, filters, and other things in Solr, so you probably know that the class attribute is responsible for specifying the class that implements the desired similarity implementation, in our case, solr.DFRSimilarityFactory. Also, if there is a need, you can specify additional parameters that configure the behavior of your chosen similarity class. In the previous example, we specified the four additional parameters of basicModel, afterEffect, normalization, and c, all of which define the DFRSimilarity behavior.

The solr.SchemaSimilarityFactory class is required to specify the similarity for each field.

Although the recipe is not about all the similarities available, I wanted to list the available ones. Note that each similarity might require and use different configuration parameters (all of them are described in the provided Javadocs). The list of currently available similarity factories are:

In addition to per-field similarity definition, you can also configure the global similarity.

<similarity class="solr.BM25SimilarityFactory"/>